A topology based approach to coordinated multicast operations

Abstract

Multicasts are a powerful means to implement coordinated operations on distributed data-sets as well as synchronized reductions of multiple computed results. In this paper we present a topology based approach to implement parallel operations on distributed data-sets as multicasts. Multicast groups are described as reusable application-specific topology classes that coordinate both the spreading of multicast messages and the collection (reduction) of the computed results. Thus global operations are controllable through applications and existing communication topologies as well as synchronization patterns can effectively be reused.